Vehicle Classification Using Electromagnetic Leakage Signals Based on Feature Fusion and Residual CNN

While the classification of electromagnetic (EM) leakage signals from various electronic devices has been explored in recent research, the use of EM leakage signals from vehicles for classification tasks remains relatively underexplored. In this study, we simulate EM leakage signals from the engines of three different types of vehicles to efficiently classify them using deep learning and machine learning models. For this evaluation, we relied on two convolutional neural network (CNN)-based deep learning models—a vanilla CNN and a residual CNN—along with two machine learning models—random forest and light gradient boosting model. In the experiments, we used features such as the Mel-frequency cepstral coefficient, Mel-spectrogram (MEL), short-time Fourier transform (STFT), and fast Fourier transform, along with the features obtained through feature fusion. Among the single features, we observed that models using STFT tended to exhibit better performance. Moreover, the deep learning models showed improved performance upon the implementation of the SpecAugment technique for data augmentation. The highest classification accuracy of 93.89% was achieved by the residual CNN model using the feature fusion of MEL and STFT in combination with SpecAugment. These findings confirm the feasibility of classifying vehicles using EM leakage signals.